QEMU disks have the werror/rerror policies: on an I/O error either
report it to the guest, or pause the VM and retry the request on
resume. A vhost-blk disk completes requests to the guest from the
kernel, so QEMU never sees the error. This series lets dm-qcow2 hold
a failed request according to the policy, tell QEMU about it, and
retry or fail it on request from userspace.
Patches 1-2 fix the existing ENOSPC delay: a use-after-free of the
event work and timer on destroy, and a suspend stalled by a request
delayed while the timer callback runs.
Patches 3-4 rename the ENOSPC delay list to a held request list and
arm the retry timer once per hold, so new failures do not postpone
the retry of the held requests forever.
Patch 5 sends an IO_ERROR change uevent when requests get held, as
QEMU cannot wait for dm events without CAP_SYS_ADMIN.
Patch 6 adds the resubmit and fail_held messages, so libvirt can
retry the held requests on VM resume and drop them on VM destroy.
Patch 7 adds the error_policy=<read>,<write> table argument with the
report, enospc and stop policies. The default keeps the current
behaviour for writes.
Patch 8 requeues held requests to the dm queue on a noflush suspend,
so a table reload or a snapshot of a VM paused on an I/O error does
not deliver the error to the guest. ploop suspends with flush for a
snapshot and a resize, so this needs the ploop.ovz change below.
Patch 9 fixes a vhost-blk hang on release when the vhost worker was
killed with I/O in flight, which a VM destroyed while its requests
are held runs into.
How the VM is paused: QEMU stops the VM with RUN_STATE_IO_ERROR on
the uevent. vhost-blk keeps running in that state, so the held
requests survive the pause and libvirt resubmits them before cont.
Any other pause stops vhost-blk, which waits for all in-flight
requests with the BQL held. So libvirt sends fail_held before virsh
suspend, save, snapshot and a migration of a running VM, and the
guest gets EIO for the requests held at that moment. vhost-blk
cannot keep them across a stop the way QEMU keeps its retry list.
QEMU blocks migration and savevm while the VM is paused on a held
request.
Userspace series, all for VSTOR-152175:
qemu.vzs, "[PATCH hci-8.1 0/2] vhost-blk: support the werror/rerror
policies":
vhost-blk: add the werror and rerror properties
vhost-blk: pause the VM on an IO_ERROR uevent from dm-qcow2
libvirt.ovz:
qemu: pass the error policy of vhost-blk disks to dm-qcow2
qemu: resubmit held vhost-blk requests before resuming the VM
qemu: fail held vhost-blk requests before killing QEMU
qemu: pass IO_ERROR uevents to vhost-blk
conf: allow error_policy for vhostblk disks
qemu: fail held vhost-blk requests before pausing the VM
qemu: fail held vhost-blk requests before migrating a running VM
ploop.ovz:
lib: add read/write error policies for dm-qcow2 devices
(passes error_policy on mount and keeps the policy of the loaded
table on every reload, snapshot, merge and replace included)
lib: add ploop_resubmit_held() and ploop_fail_held()
ploop: add the -e error policy option to ploop mount
lib: suspend unpartitioned dm-qcow2 devices with noflush
and tests for both
v3:
- cover letter: describe how userspace pauses a VM with held
requests, list the qemu.vzs, libvirt.ovz and ploop.ovz changes the
series needs, note the ploop dependency of patch 8
v2:
- patches 1, 2, 4, 9: add Fixes tags
- patch 5: state that held ENOSPC reads are reported too, and that
every dm-qcow2 device whose storage stays full now gets a uevent on
each retry
- patch 7: drop the misleading "whatever the error" from the message
- patch 8: note that ploop needs the ploop.ovz change "lib: suspend
unpartitioned dm-qcow2 devices with noflush" to get the requeue on
a snapshot or resize
https://virtuozzo.atlassian.net/browse/VSTOR-152175
Alexander Ivanov (9):
drivers/md/dm-qcow2: cancel the event work and ENOSPC timer in the
destructor
drivers/md/dm-qcow2: always resubmit delayed ENOSPC requests in
presuspend
drivers/md/dm-qcow2: rename the ENOSPC delay list to a held request
list
drivers/md/dm-qcow2: arm the hold timer only for the first held
request
drivers/md/dm-qcow2: report held requests with an IO_ERROR uevent
drivers/md/dm-qcow2: add the resubmit and fail_held messages
drivers/md/dm-qcow2: add the error_policy table argument
drivers/md/dm-qcow2: requeue held requests on a noflush suspend
drivers/vhost/blk: fix a hang on release after the vhost worker is
killed
drivers/md/dm-push-backup.c | 2 +-
drivers/md/dm-qcow2-cmd.c | 32 +++++++++-
drivers/md/dm-qcow2-map.c | 110 +++++++++++++++++++++++++++++------
drivers/md/dm-qcow2-target.c | 104 ++++++++++++++++++++++++++++-----
drivers/md/dm-qcow2.h | 26 +++++++--
drivers/md/dm-rq.c | 4 +-
drivers/md/dm-rq.h | 2 +-
drivers/vhost/blk.c | 52 ++++++++++++-----
8 files changed, 274 insertions(+), 58 deletions(-)
--
2.53.0
_______________________________________________
Devel mailing list
[email protected]
https://lists.openvz.org/mailman/listinfo/devel